“Leveraging Human Brain Activity to Improve Object Classification” with Professors David Cox and Walter Scheirer.*
Spring 2014 - Present
As an undergraduate research assistant for Professor Cox, I have been conducting object detection research in biologically inspired machine vision algorithms.
Abstract: How can computers think and act like humans? In the past few decades, major breakthroughs have been made in solving challenging yet clearly-defined tasks, such as winning a chess game or Jeopardy. Yet, computers still perform poorly on intuitive tasks that humans do subconsciously, such as recognizing different kinds of fruits. Currently, most algorithms attempting these vision tasks differ drastically from how the human brain tackles these problems. My thesis investigates how we can improve machine vision algorithms by designing them to better mimic how humans approach these tasks. Specifically, I am examining how human brain activity from functional magnetic resonating images (fMRI) can be leveraged to improve object classification. Inspired by the idea that humans learn in a graduated manner and generalize their experiences with easy problems to tackle harder ones, I am developing techniques first to learn the difficulty of detecting an object in an image given auxiliary fMRI data and second to modify vision algorithms to incorporate these difficulty measures.
Senior Thesis: http://bit.ly/ruth_fong_senior_thesis.
Cox Lab Meeting Presentation: http://bit.ly/rfong_fmri_annotations_lab_talk.
“Ensuring Privacy for Genomics Data With K Disease Categories” with Louis Li.
Genome-wide association studies (GWASs) are used to provide aggregate statistics about the influence of single-nucleotide polymorphisms (SNPs), which are units of genetic data, on disease outcome. Previous research (Uhler '13) demonstrates a differentially private way to release three kinds of statistics – averaged minor allele frequencies (MAFs), χ-squared statistics, and p-values – for a specified number of SNPs using the Laplace mechanism with the constraint that the disease outcome is binary – typically healthy or sick. We extended some of Uhler’s theoretical results to provide differential privacy for GWAS data with K disease outcomes and experimentally analyzed the utility of our generalized mechanisms. These extensions are significant because they allow for the private release of more forms of GWAS data, particularly that of diseases with more complex disease categories or several levels of severity.
Final Project Paper: http://bit.ly/rfong_cs227r_final_paper.
“Human-Animal Look-a-likes: Exploring measures of similar across object categories.”
For the final project of a graduate-level computer vision course, I investigated methods for matching images of human faces with images of similar-looking animal faces. I generated this research idea with inspiration from a humorous British tabloid article that juxtaposed images of Benedict Cumberbatch with photos of otters that looked uncannily like the famous British actor. How do humans recognize similarities across different kinds of objects? While significant research in the computer vision community has been conducted on facial recognition and object recognition, similarities across object categories have been largely unexplored. I compared how three different similarity metrics – Euclidean distance, Mahalanobis distance, and One-Shot similarity scores – selected similar-looking pairings between human and animal faces when using Histogram of Oriented Gradient descriptors to describe images. I developed a framework for comparing the similarity of two given images using one of these metrics. I designed and ran an online experiment that solicited participants to select the best pairings.
Final Project Paper: bit.ly/rfong_cs283_final_paper.
“Word Sense Disambiguation” with Louis Li.
For the final project of a computational linguistics course, my partner and I analyzed algorithms that infer the meaning of a word from surrounding context words in a sentence. We compared Lesk’s seminal dictionary-based algorithm (Lesk 1986) and Mihalcea’s graph-based sequence data labeling method (Mihalcea 2005). I implemented the Lesk algorithm to use the Oxford dictionary and the WordNet knowledge source.
Final Project Slides: bit.ly/rfong_cs187_final_slides.
“Imperial Search: An Analysis of the Impact of British Imperialism in the 20th Century.”
For the final project of a British history course, I developed an interactive tool to compare the relative influence of two topics and learned the most prominent topics in each decade of the 20th Century British parliament using a Latent Dirichlet Allocation topic model. Interested in quantifying the impact of British imperialism in Parliament after reading a claim (Porter 2004) that the empire had an insignificant effect on Britons at home because issues of imperialism only arose in Parliament 10-15% of the time, I generated this research idea.
*Working towards publication.